Safety right now isn't about keeping today's model from hacking NORAD. It's about learning the techniques and approaches to try and have at least a halfway dec

Safety right now isn't about keeping today's model from hacking NORAD. It's about learning the techniques and approaches to try and have at least a halfway decent approach to safety by the time models capable of autonomously hacking NORAD exist. (It's also just a very fun problem space in general.)